Tether has released a medical AI model called QVAC MedPsy that runs entirely on a smartphone and outperforms Google's larger MedGemma-27B in real-world scenarios. The model uses three times fewer compute resources and is 16 times smaller than the Google version it beats, according to the company.
Performance vs. Size
Benchmarks show QVAC MedPsy scoring higher than MedGemma-27B on tasks that mimic actual clinical use. While Google's model requires substantial server-level processing, Tether's model fits on a mobile device. The efficiency gain comes from the model's compact architecture — 16 times smaller by parameter count — yet it still delivers superior results.
That size difference translates directly to compute needs. QVAC MedPsy consumes roughly one-third the processing power of MedGemma-27B, making it viable for on-device inference without internet access.
Why Efficiency Matters for Medical AI
Running AI on smartphones opens up use cases in remote clinics and low-resource settings where a stable server connection isn't guaranteed. A model that fits on a device can analyze patient data locally, cutting latency and privacy risks. Tether hasn't disclosed specific deployment plans, but the model's performance suggests it could handle psychiatric or psychological screening — the “Psy” in QVAC MedPsy likely refers to psychology or psychiatry.
The medical field has been cautious about cloud-only AI due to data sensitivity. An on-device model that still outperforms big competitors could speed adoption in hospitals and mobile health apps.
The Tech Behind It
Tether, best known for its stablecoin USDT, has been quietly building in AI. QVAC MedPsy is one of its first medical models. The company has not published detailed network architecture, but the model's ability to run on a smartphone while beating a much larger model indicates aggressive optimization in training or quantization.
Google's MedGemma-27B, by contrast, is designed for cloud or high-end hardware. Tether's approach flips that assumption, showing that smaller can be stronger in domain-specific tasks.
No launch date or partnership has been announced. For now, the benchmark results stand as a proof point that medical AI does not always need big iron to deliver big results.
